Motif Simplification: Improving Network Visualization Readability with Fan, Connector, and Clique Glyphs Cody Dunne, Ben Shneiderman Department of Computer Science and Human-Computer Interaction Lab University of Maryland, College Park, MD 20742 fcdunne, [email protected] (a) 65% (b) 65% simplified (c) 70% (d) 70% simplified Figure 1: U.S. Senate 2007 co-voting network at 65% and 70% agreement cutoffs, simplified using clique motif glyphs. Key features are visible, such as the moderate Republican clique around McCain with “wildcards” at the periphery. ABSTRACT Author Keywords Analyzing networks involves understanding the complex re- Motif simplification; network visualization; graph drawing; lationships between entities, as well as any attributes they node-link diagram; visual analytics. may have. The widely used node-link diagrams excel at this task, but many are difficult to extract meaning from be- ACM Classification Keywords cause of the inherent complexity of the relationships and lim- H.5.2. User Interfaces (D.2.2, H.1.2, I.3.6) ited screen space. To help address this problem we introduce a technique called motif simplification, in which common INTRODUCTION patterns of nodes and links are replaced with compact and Networks of entities and their ties have long been common meaningful glyphs. Well-designed glyphs have several ben- data structures in computer science, but have only recently ex- efits: they (1) require less screen space and layout effort, (2) ploded into popular culture with publishers like the New York are easier to understand in the context of the network, (3) Times including elaborate and interesting networks with their can reveal otherwise hidden relationships, and (4) preserve articles. Online communities like Facebook, MySpace, Twit- as much underlying information as possible. We tackle three ter, Flickr, and mailing lists (to name only a handful) enjoyed frequently occurring and high-payoff motifs: fans of nodes enormous growth over the last few years and provide incred- with a single neighbor, connectors that link a set of anchor ibly rich datasets of interpersonal relationships, which social nodes, and cliques of completely connected nodes. We con- scientists are now fervently exploring. Networks have also tribute design guidelines for motif glyphs; example glyphs for found applications in such diverse disciplines as bioinformat- the fan, connector, and clique motifs; algorithms for detect- ics, scientometrics, urban planning, politics, and archeology. ing these motifs; a free and open source reference implemen- Analysis of network data requires understanding clusters, tation; and results from a controlled study of 36 participants connectivity, and centrality. Statistical analysis and conven- that demonstrates the effectiveness of motif simplification. tional visualization tools like bar and pie charts are often inad- equate when faced with these varied and oftentimes immense datasets. visualcomplexity.com provides almost 800 network Permission to make digital or hard copies of all or part of this work for visualizations, but most are variations of node-link diagrams, personal or classroom use is granted without fee provided that copies are where nodes represent entities and the links or edges indicate not made or distributed for profit or commercial advantage and that copies ties connecting them. Node-link diagrams only recently be- bear this notice and the full citation on the first page. To copy otherwise, or came widely available but have already been put to great ef- republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. fect, such as detecting social roles in online newsgroups [32] CHI 2013, April 27–May 2, 2013, Paris, France. or studying U.S. political blog ties during an election [2]. Copyright 2013 ACM 978-1-4503-1899-0/13/04...$15.00. However, there is a huge array of possible layouts of the nodes and links in any given network, many of which can be misleading or incomprehensible. Network visualizations are only useful to the degree they “effectively convey infor- mation to the people that use them” [3]. In fact, the spatial layout of a node-link diagram can have a profound impact on the detection of communities in the network and the perceived Figure 2: From left to right: fan, connector, and clique motifs. importance of actors [23]. Significant thought must be given to proper visualizations so that analysts will be able to un- derstand and effectively communicate data like clusters, the paths between them, and the importance of individual actors. RELATED WORK Network analysis tools generally use node-link diagrams, As manual layout of nodes in the node-link diagram is incred- as in NodeXL [27], or matrix representations like Matrix- ibly time consuming to do well, a lot of effort has been put Explorer [13]. Both node-link and matrix diagrams show into developing automated network layout algorithms and fil- the topology of small networks well but can be unreadable tering tools. As the optimization of many readability metrics with a few thousand nodes. We can reduce the visualiza- is NP-hard [3], layout algorithms often use heuristics that pro- tion complexity by showing an aggregate version of the net- duce suboptimal visualizations quickly. However, the results work, based on any number of criteria. NetLens [17] groups of applying a layout algorithm can vary greatly depending on nodes by their attributes and can pivot between connected the size and topology of the network, and the layout gener- groups of two different types, while PivotGraph [31] uses ated is highly dependent on the algorithm used. We believe attribute groupings but shows ties between aggregates using that state of the art layout algorithms alone are insufficient to arcs. GraphTrail [5] combines these approaches with familiar consistently produce understandable network visualizations. charts, arc diagrams, and a many-to-many pivot between sev- One way forward is the use of aggregation, specifically by ag- eral node types. However, these approaches focus on attribute gregating common network structures or subnetworks called comparisons at the expense of showing topology. motifs. Large, complex network visualizations often have Alternatively, we can use a hierarchical topologic clustering motifs repeated throughout because of either the network to show a network of meta-nodes like ASK-GraphView [1] structure or how the data was collected. Regardless of their and van Ham & van Wijk [30]. Rather than letting meta- cause, some frequently occurring motifs contain little infor- nodes overlap, van Ham & van Wijk used semantic fisheye mation compared to the space they occupy in the visualiza- views to show clusters as merging spheres. Other approaches tion. Existing tools may highlight certain motifs, allow users to creating overview networks include graph summarization to filter them out manually, or replace them with meta-nodes. [25] and aggregating nodes by shared neighbor sets [21]. [21] We improve on these approaches with motif simplification, also provide a topologic clustering tool, and a level of detail in which network motifs are automatically replaced with option to split meta-nodes apart to better see the underlying compact, representative glyphs. Well-designed glyphs have topology. ManyNets [7] takes a different tack, showing sta- several benefits: they (1) require less screen space and lay- tistical comparisons of a network partitioned by topology, at- out effort, (2) are easier to understand in the context of tributes, or time. In each of these techniques it can be difficult the network, (3) can reveal otherwise hidden relationships, to understand the topology of the individual aggregates, often and (4) preserve as much underlying information as possi- because of the ambiguous nature of clustering algorithms. ble. In this paper we discuss three high-payoff motifs that Instead of clustering, we can use a metric for node importance plague network analysts, shown in Fig.2: fans, connectors, to filter to an important subset. Skeletal images [15] high- and cliques. We contribute the design of representative and lights high-metric nodes, and replaces filtered trees with tri- combinable glyphs for these motifs, algorithms for detect- angles that take the same space. Tsigkas et al. [29] similarly ing them, and a supporting task-based controlled study with filtered a security network of events and features on a domain- 36 participants. These techniques are all implemented and specific metric, while including a way to aggregate the events made publicly available as part of the free and open source joining a subset of features into meta-edges. However, the NodeXL network analysis tool [27], which is available from aggregation is limited to ties between two feature types and nodexl.codeplex.com. obscures the number of connecting nodes and edges. Specifically, the contributions of this paper are: Our approach is to aggregate the network by the frequently • A technique for simplifying node-link diagrams by replac- occurring motifs it contains. While the fan, connector and ing common network motifs with representative glyphs, clique motifs we target are quite prominent in social network • A set of design guidelines for these glyphs to show the mo- datasets, there are many other motifs of interest, especially tif contents and underlying attributes, for biologists. Motif census (counting the kinds of motifs) • The design of glyphs for fans, connectors, and cliques, and analysis is used extensively to analyze the behavior of • Algorithms for detecting these three motifs, complex biologic networks, looking for repeated patterns that indicate underlying processes. For example, Milo et al. [24] • A supporting task-based study with 36 participants, and find motifs that appear more frequently than expected in ran- • A free and open source implementation as part of NodeXL. dom networks, and provide a chart of small motif frequency. Knowledge of the motifs present in a network can help predict behavior and the “structural signatures” of individual entities [32], but visualizing these motifs effectively is challenging. Huang et al.